t0kt0

writing about web craft, CSS, design engineering 🌈 by & for diversity in tech

12 August 2026

aiwritingclaude

Teaching Claude to sound like me

How I turned 937 voice dictations into a voice spec, and how my own reviews fixed the skills that use it.

From May 2026 to now.

I dictate a lot. Messages, PR comments, and a lot of instructions to coding agents go through Wispr Flow before they become text. At some point I noticed something annoying. The text I dictated sounded like me. The text an AI wrote for me sounded like everybody.

You know the voice. "I hope this message finds you well." "Furthermore." A perfectly balanced sentence with an em dash in the middle. Nothing is wrong with it. It just isn't mine, and people can tell.

937 dictations

Wispr Flow keeps a local history of everything you dictate, in a SQLite database on your machine. In May 2026 I had an agent read that history and tell me how I actually talk. Not how I think I talk.

There were 937 dictations: 637 in English, 169 in Spanish, and 131 it couldn't place. The English ones came mostly from the terminal and work tools. The Spanish ones came mostly from WhatsApp.

The result was a voice spec, one Markdown file in my notes repo. Some of it I could have guessed. Some of it was a bit uncomfortable to read.

  • In English I lead with the action: "Make sure...", "Let's...", "Update...". I correct myself mid-sentence with "actually" or "also", and I say "I don't like it" when I don't like it.
  • In Spanish I'm a different person. Warmer, looser, more narrative. "Igual", "pero bueno", "nada", "si te soy honesta". The sentences wander a bit before they land.
  • In both, I can't stand filler. No fancy transitions, no fake warmth, no em dashes.

That last one became the first hard rule. The spec ends with a list of non-negotiables for any model that only reads one section, and number one is "Do not use em dashes."

From a file to skills

A spec in a file only helps if something reads it. So I wrapped it in two Claude Code skills.

my-voice takes a dictated transcript and turns it into a message: Slack, a PR description, WhatsApp, email, website copy. It strips the "um" and the false starts, keeps the phrases I actually chose, and applies the rules for that format. On top of the spec it adds one rule that comes from my own pet peeve: American English only. Color, not colour.

draft is the entry point for everything else: bullets, a one-line ask, a thread to answer, a git diff. It hands the voice back to my-voice and adds a layer for formats, like how I write a PR description, a weekly project update or a product announcement.

The order matters. The spec is the base, and it changes. The skills read it every time instead of keeping their own copy. When I learn something that should apply everywhere, it goes into the spec, not into a fourth copy of the same rules.

The baseline won

In August 2026 I ran evals on draft: the same prompts with the skill and without it. I read the outputs side by side and wrote down what I thought of each one.

The Spanish WhatsApp message with the skill: "It's perfect." The PR description: "Awesome. Looks great."

Then the Slack update. With the skill, my note was "It's a bit too dry." Without the skill: "Way better. I think this is nicer."

That one stung a bit, but it made sense once I looked at the data again. Most of my English dictations were instructions to agents and tools. So the spec had learned how I talk when I'm steering work: short, imperative, clipped. That's not how I write to my team. To people I say "Hi team, quick update", I write full sentences, and I end with "Thanks!". The spec had even warned me. Formal English was on its "less certain" list, because the sample was mostly work instructions.

So I added a calibration section to draft, written from my own reviews:

  • Tight means no filler, not no warmth. Open with a greeting, close with a thanks.
  • In informal Spanish, skip the opening "¿" and "¡". They're not used in informal writing.
  • In Slack lists, use dashes, not bullet characters.
  • In PR notes, a few short bullets beat one dense paragraph.

None of these are deep. They are the exact things I used to fix by hand every time, and now I don't have to.

Never send

The most important rule didn't come from the voice at all. In that same eval round, one of the test drafts actually got sent to Slack. The text was fine. The problem was that it left my hands before I read it.

So draft now has a contract at the top: it writes text, and it never sends it. Not to Slack, not as a Gmail draft, not as a PR. Not even when a connector could do it, or when I say "draft and send". It also never asks "want me to send it?". Sending is my call, after I read it.

That's the part I'd copy first if you build something like this. A voice that's 90% right is great. A voice that's 90% right and posts on your behalf is a problem.

Where it is now

The skills live in my notes repo next to the spec. A small script packages them with a copy of the spec, so they also work on claude.ai, where my local files don't exist. When I correct a draft for something, the fix goes into the calibration list.

Every correction makes the next draft sound a bit more like me. Whether this one passes the test, I'll let you decide.